AI labs are handing more of their own research over to their AI systems, and the people building it are openly worried it could go badly wrong.

• Anthropic now runs a "wet lab" where its AI directs real biology experiments, aimed at speeding up drug and disease research.
• AI is starting to improve itself: Anthropic says its own model leads the work on a quarter of its research tasks, and OpenAI claims its model cracked more than a hundred long-open math problems.
• Kevin Roose, who spent two years reporting inside these companies, says the staff privately think the odds of catastrophe are worse than what they admit publicly.
• There are almost no laws written for this — product liability rules only cover products actually sold to the public, so an internal model that causes harm falls through the cracks.
• Both Anthropic and OpenAI are heading for IPOs; Palantir's Alex Karp thinks the liability is so large they will end up nationalized instead, though their revenue is growing faster than any company ever.

Outlook: Anthropic's listing is the nearer of the two, and pressure is building for a ban on AI that improves itself before either company goes public.